Clinical Report: Machine Learning in Liver CT Imaging – Efficacy and Applications
Overview
This systematic review analyzed 191 studies on machine learning (ML) applications in liver CT imaging, focusing on segmentation, detection, and classification tasks. ML models, especially deep learning, demonstrated high performance in liver and lesion segmentation, with some achieving near-expert accuracy, though challenges remain in small lesion detection and clinical applicability.
Background
Machine learning is increasingly recognized as a reliable tool in medical imaging, with over 220 ML-based medical devices approved in the USA and Europe. CT imaging is critical for diagnosing and monitoring liver diseases, and ML applications have been explored for liver and lesion segmentation, tissue characterization, and disease prediction. Despite promising results, comprehensive evaluation of ML performance and clinical relevance in liver CT imaging has been limited, prompting this systematic review.
Data Highlights
Study Aim
Number of Studies
Highest Reported DICE Score
Lowest Reported DICE Score
Liver Segmentation
84
0.9851
0.75
Lesion Segmentation
60
Not specified
Not specified
Key Findings
ML was applied primarily for liver segmentation (84 studies), lesion segmentation (60 studies), lesion detection, classification, and other tasks.
Deep learning methods dominated and showed high accuracy, with liver segmentation DICE scores ranging from 0.75 to 0.9851.
External validation was increasingly used to improve generalizability, but only 11 studies compared ML performance directly with human experts.
Most studies relied on publicly available datasets, such as LiTS 2017, highlighting challenges in obtaining labeled medical data.
Lesion segmentation models performed well for lesions larger than 2 cm but struggled with lesions smaller than 1 cm, similar to clinical limitations.
Reporting transparency varied; only a minority of studies provided confidence intervals or standard errors for performance metrics.
Clinical Implications
ML-based tools show promise for enhancing liver CT imaging analysis, particularly in segmentation tasks, potentially aiding radiologists in diagnosis and treatment planning. However, limitations in small lesion detection and variability in study quality suggest cautious integration into clinical workflows. Further validation against expert performance and standardized reporting are needed to support clinical adoption.
Conclusion
Machine learning, especially deep learning, demonstrates high efficacy in liver CT imaging tasks, with potential clinical benefits. Continued research focusing on clinical validation and addressing current limitations will be essential for broader clinical implementation.
References
Systematic Review Authors 2022 -- Evaluation of Machine Learning Efficacy and Clinical Relevance in Liver CT Imaging: A Systematic Review
